LBP-based periocular recognition on challenging face datasets

EURASIP Journal on Image and Video Processing

Table 2 Rank-1 accuracies ND Twins effect of periocular image size and image scale

Pyramid levels ( P), image size ( S × S)	Approach	Left (%)	Right (%)	α-Fusion (%)
P = 4,S = 256	H-3P-LBP	96.60	98.62	98.11
	LBP	52.83	56.23	66.79
P = 4,S = 128	H-3P-LBP	96.60	98.62	98.11
	LBP	52.83	56.23	67.30
P = 3,S = 256	H-3P-LBP	88.72	84.85	90.49
	LBP	55.56	51.01	65.32
P = 3,S = 128	H-3P-LBP	97.42	98.33	98.03
	LBP	53.18	52.73	70.15
P = 3,S = 100	H-3P-LBP	86.95	85.52	89.8
	LBP	56.31	51.94	64.90
P = 1,S = 256	3P-LBP	96.98	96.60	98.11
P = 1,S = 128	3P-LBP	96.98	96.60	98.11

The accuracies for LBP are achieved using the periocular image from the finest scale of the image pyramid. The 3P-LBP was achieved with P = 1 (the last two rows of the table). Values in italics are the best achieved results.